Automatic music classification into genres

Abstract

Musical genres are categorical labels created by humans to characterize pieces of music. Although music genres are inexact and can often be quite arbitrary and controversial, it is believed that certain song characteristics like instrumentation, rhythmic structure, and harmonic content of the music are related to the genre. In this paper, the task of automatic music genre classification is explored. Multiple features based on timbral texture, rhythmic content and pitch content are extracted from a single music piece and used to train different classifiers for genre prediction. The experiments were performed using features extracted from one or two 30 second segments from each song. For the classification, two different architectures flat and hierarchical classification and three different classifiers (kNN, MLP and SVM) were tried. The experiments were performed on the full feature set (316 features) and on a PCA reduced feature set. The testing speed of the classifiers was also measured.The experiments carried out on a large dataset containing more than 1700 music samples from ten different music genres have shown accuracy of 69.1% for the flat classification architecture (utilizing one against all SVM based classifiers). The accuracy obtained using the hierarchical classification architecture was slightly lower 68.8%, but four times faster than the flat architecture.